Research And Optimization Of Pedestrian Detection Based On YOLOv3 Algorithm

Pedestrian detection,as a sub-problem of object detection,is one of the popular research directions in the field of computer vision.It has huge application value in the fields of automatic driving,video surveillance,security,etc.Many companies and universities have conducted research on pedestrian detection.Because pedestrian targets are affected by many factors,pedestrian detection still has many difficulties and challenges:(1)Pedestrian detection requires high real-time performance to meet application requirements;(2)Small-sized pedestrians(height is less than 60 pixels)are difficult to detect;(3)Pedestrians often have overlapping and shadowing.Based on the above problems,this thesis focuses on the improved pedestrian detection algorithm based on YOLOv3 network.The improved YOLOv3 algorithm can detect pedestrians of different heights and sizes at different scale feature layers more accurately,and can meet the requirements of high real-time performance.The main work of this thesis includes the following content:(1)Improved YOLOv3 network based on multi-scale feature fusion.In order to make full use of the spatial information of the underlying features and the semantic information of the high-level features,when feature fusion is performed at three scales,the two feature layers except the current scale feature layer are first resized to the current scale.After the feature layer has the same size,through the 1 x 1 convolution kernel,output 3 weight parameters of the current scale,and then use this to perform feature fusion.On the BDD100 K pedestrian data set,the detection effect of pedestrians at different heights based on YOLOv3 network and using adaptive feature fusion and K-means ++ clustering algorithm is verified.(2)Optimization of YOLOv3 bounding box regression algorithm and NMS algorithm.In order to obtain a more accurate prediction box position,the boundary box regression loss function Center?dis-IoU based on the center point distance is used to replace the MSE loss of the coordinate loss in the YOLOv3 loss function,then Soft-NMS and Center?dis-IoU are used to improve the NMS algorithm to reduce the missed detection rate.Then on the Pascal VOC2007 data set,the loss part of the coordinates in the loss function was replaced with a new bounding box regression loss function to verify its improvement on the training convergence speed and the accuracy of the positioning of the bounding box and using Soft-NMS,Center?dis-IoU Validate optimization for removing duplicate prediction boxes(3)Pedestrian detection experiment and optimization based on improved YOLOv3 algorithm.Combine these improvements,and use pre-trained model weights based on mix-up strategy training to perform multi-scale training on the BDD100 K data set.The experimental results show that the detection accuracy of the model is significantly improved,the missed detection rate is greatly reduced,and it has high application value.